Segmentation of Cardiovascular Structures from Screening Non-contrast CT Calcium Score Images Using the Tuned CardioNC-Segmentator on Virtual CTA Images

In this work, we outline the development of a deep learning-based segmentation network aimed at automatically segmenting heart chambers using virtual CT angiography (VCT A) data. Inspired by the structure and methodology of TotalSegmentator, our approach leverages nnU-Net architecture, optimized for the precise delineation of anatomical structures in volumetric medical images. The model is trained on a dataset (N=51) of non-contrast CT calcium score (CTCS) scans, annotated with ground truth segmentations of the heart's atria and ventricles. The integration of virtual CT angiography as a pre-processing step enhances the visibility of cardiac structures, such as large blood pools and adipose and muscle tissues, aiding the model's ability to discern complex boundary regions. Our results demonstrate the network's ability to achieve high agreement in segmenting the left and right atria and ventricles, showing potential for automating the diagnosis and treatment planning of cardiac conditions. The bias for the myocardium and heart chambers averaged 5± 14.2 %. The developed segmentation tool is poised to improve workflow efficiency in clinical settings by providing reliable, reproducible heart chamber segmentations from virtual CTA scans.